Research on Theory and Technique of Ultra-wideband SAR Shallow Buried Targets Imaging and Detection G&
m~W
Jin Tian GRh 430V [
ABSTRACT % 'OY
There are up to 110 million landmines in many countries all over the world, which thread people’s safe and obstruct economic development seriously. In addition, it will cost 300 to 1000 dollars to eliminate a 30-dollar landmine. At the current investment and technology, it will take about 1400 years to remove all landmines. Therefore, an efficient landmine detection instrument is urgently needed. Air- or vehicle-borne ultra-wideband synthetic aperture radar (SAR) can perform quick detection of shallow buried targets over large areas, which overcomes the low-efficiency and low-safety shortcomings of those traditional landmine detection measurements, and has became a new tread of landmine detection technique. With the support of the defense pre-research project and the weapon equipment demonstration project, this dissertation has comprehensively studied the basic theory of ultra-wideband SAR imaging and detection and proposed the imaging and detection integrated procedure for shallow buried targets and the time-frequency representation based realization method. In the imaging and detection integrated framework, shallow buried object refraction and dispersion compensation, radio frequency interference (RFI) suppression, speckle noise suppression, shallow target feature extraction and discriminator design have been investigated and some useful solutions have been obtained, which place a great role in the following equipment research project. The major work and innovations are introduced as follows: pZz\o
1) The problems in the traditional ultra-wideband SAR imaging and detection procedure have been comprehensively studied. Based on the large relative bandwidth and wide beamwidth characteristic of ultra-wideband SAR, the “imaging and detection integrated framework” concept and its time-frequency representation based realization method are proposed. The framework employs the frequency and aspect angle information of ultra-wideband SAR target scattering to improve the processing performance, which includes two parts: “detection oriented imaging” and “imaging based detection”. The “detection oriented imaging” mainly focuses on how to extract the frequency and aspect angle information of target scattering from the received echo to improve prescreening performance when obtaining high-resolution SAR images; The “imaging based detection”, operating on the extracted suspected targets by prescreening, mainly focuses on how to extraction the frequency and aspect angle information of target scattering from formed images to eliminate clutter and improve discrimination performance. The relationship of imaging and detection is often neglected in the traditional imaging and detection procedure and thus cannot solve the contradiction between the image resolution and the frequency and aspect angle information extraction precision, which limits the improvement of the final detection performance. In this dissertation, based on the developed target echo and imaging models of ultra-wideband SAR, a detection oriented time-frequency representation image formation (TFRIF) is proposed, and its image domain form can also solve the problem of “imaging based detection” for shallow buried targets. The TFRIF based imaging and detection integrated realization method can extraction the frequency and aspect angle information of target scattering accurately without sacrifice of resolution, which not only improves the prescreening and discrimination performance but also improve the whole processing efficiency of the imaging and detection procedure. OE*Y%*b
2) Considering the problem of shallow buried object echo distortion caused by the multi-layered propagation media of air and soil, the refraction and dispersion effects compensation for shallow buried object imaging is studied. With the priori-knowledge of buried depth, incident angle, and so on, two shallow buried image formation with the echo domain refraction and dispersion compensation (EDRDC), modified wavefront reconstruction (MWR) and subsurface back-projection (SBP), are proposed. Compared with the refraction point calculation based time domain algorithms and the phase shift based frequency domain algorithms, MWR and SBP not only have better computational efficiency but also consider the frequency varying characteristic of the soil relative permittivity to yield better focusing and locating performance. But the priori-knowledge of buried depth, incident angle, and so on cannot be obtained in practical applications, and thus the shallow buried object focusing and locating method with the image domain refraction and dispersion compensation (IDRDC) in the imaging and detection framework is proposed. The IDRDC method estimates the buried depth on the maximum azimuth compression amplification criterion and its compensation factor considers not only the soil dispersion characteristic but also the incident angle varying with radar at different azimuth positions. Therefore, the IDRDC method has better focusing and locating performance than the EDRDC method. The IDRDC method operates on each object in the image domain and thus can focus and locate multiple shallow buried objects in different buried depths and soil environments, which fulfill the practical requirement of large areas detection for air- and vehicle-borne ultra-wideband SAR. )$ M2+_c
3) The practical RFI and speckle noise suppression techniques for ultra-wideband SAR are proposed. In the aspect of RFI suppression, considering the problem of TV, broadcast and communication signal in the ultra-wideband SAR operation frequency degrading image quality, a novel Wiener filter construction method for RFI suppression is proposed. Because RFI signals have the non-stationarity characteristic, the traditional Wiener filter construction method need record real-time RFI. The novel method is based on the region of support difference of target echo and RFI in the two-dimensional frequency domain to estimate the RFI frequency spectrum, and thus can employ the radar received signal, including target echo and RFI signal, to construct the Wiener filter, directly, which reduces the system complexity and ensures the good RFI suppression performance. In the aspect of speckle noise suppression, considering the operation characteristic of vehicle-borne forward-looking ultra-wideband SAR that it continuously obtain several images of different depression angles on the same area when moving ahead, the multi-look technique is adopted to suppression speckle noise. In order to improve the registration efficiency of the several images of different depression angles in the multi-look processing, the ground-plane focusing back-projector image formation and its associated refraction and dispersion compensation technique are proposed. The images in the slant-plane have spatial varying distortion and thus need complex registration operation, but only shift exists among those ground-plane images of different depression angles. 6= :s3I^
4) On the two typical shallow buried targets, metallic landmine and unexploded ordnance, the shallow buried target feature extraction technique is studied in the imaging and detection integrated framework. For metallic landmines, the electromagnetic model of a shallow buried metallic landmine is built via the physical optics method to analyze the double-peak feature of the metallic landmine quantitatively and then the metallic landmine double-peak feature enhancement algorithm in the image domain is proposed. Furthermore, the space-wavenumber distribution (SWD) based method to estimate the four dimensional metallic landmine scattering function of slant range, azimuth, frequency and aspect angle and its associated feature selection method are proposed, which obtain the feature vector with the double-peak and the aspect-invariance features. For unexploded ordnances, the four dimensional unexploded ordnance scattering function of slant range, azimuth, frequency and aspect angle is firstly estimated via the SWD. Secondly, the amplitude information of the multi-aspect feature in different frequencies is extracted, and the spatial distribution information of multi-aspect feature in different frequencies is quantitative described using the Hu moment invariants further. Compared with the subband-subaperture technique used in the traditional imaging and detection procedure, the metallic landmine and unexploded ordnance feature extraction methods in the imaging and detection integrated framework can obtain the frequency and aspect angle information while maintain high resolution, and thus can obtain more efficient feature vectors. YS:p(jtd
5) The fuzzy hypersphere support vector machine (FHS-SVM) is proposed for shallow buried target discrimination, and its two aspects, hyperparameter optimization and kernel function selection, are comprehensively studied. Shallow buried target discrimination have several characteristics: a small training sample set, without typical clutter samples, different misclassification risks for shallow buried target and clutter, and buried environment diversity. According to the above characteristics of shallow buried target discrimination, the hyperplane SVM, which has achieved the best classification results for many classification problems in different fields by now, has been modified to yield the FHS-SVM, which uses the hypersphere in the kernel space to separate shallow buried target and clutter. The FHS-SVM based on the structural risk minimization criterion can solve the small sample learning problem, which can get a good shallow buried target and clutter classification performance with only shallow buried target training samples to obtain the parameters of hypersphere. Furthermore, the factors of misclassification risk and bury environment diversity are combined into the discriminator study procedure using the fuzzy membership of training samples, which improve the practical value of the FHS-SVM in shallow buried target discrimination. In the aspect of hyperparameter optimization, the equality between the FHS-SVM and the first level Bayesian inference is proved, and the evidence framework based hyperparameter optimization method for the Gaussian kernel FHS-SVM is proposed, which reduces the total misclassification risk of the detection result and improve the metallic landmine and unexploded ordnance discrimination performance. The evidence framework based hyperparameter optimization method can ensure good optimization performance and has better computational efficiency than those exhaustive search methods such as the margin distribution analysis, error upper bound approximation, and so on. In the aspect of the kernel function selection, the Gaussian kernel is replaced by the hidden Markov model kernel, which describes the multi-aspect characteristic of unexploded ordnance, to improve the FHS-SVM unexploded ordnance discrimination performance further. The employment of hidden Markov model kernel FHS-SVM to the unexploded ordnance discrimination combines the multi-aspect characteristic of unexploded ordnance scattering into the discriminator design, which shows the thought of “imaging based detection” that use the frequency and aspect angle information of target scattering to improve the discrimination performance. <*(^QOM
h0 $Y;=YA
Key words: ultra-wideband, synthetic aperture radar, ground penetrating, subsurface imaging, shallow buried target detection, feature extraction, discriminator design
